Projects
Major research programs, selected studies, software, and public knowledge resources. Papers are on the selected publications page; full detail is in the CV.
Major research programs
Long-running efforts that connect scientific-agent systems with action, trajectory, and policy representation.
-
01 · Agentic system · Feb 2026 – present
SIDERIUS
A closed-loop agentic system for autonomous research: it proposes, implements, and evaluates machine-learning experiments under real compute budgets. Work spans multi-agent workflow design, online resource control, and reliable long-horizon execution.
- Sole architect and developer
- Paper (arXiv) ↗
-
02 · Methodology and infrastructure · Jul 2026 – present
SciTra
A methodology, and the infrastructure that implements it, for representing what a scientific agent actually did: how a decision trajectory is recorded, and how the action space is defined, so that runs from different agents, tasks, and fields can be read on the same terms. Resource use is one of the quantities that representation makes measurable. None of it is tied to a single benchmark or a single field. Built with approximately 100 scientists across 20+ institutions in collaboration with BenchFlow.
- Founder and Coordinator
- scitra.org ↗
- BenchFlow ↗
Websites and knowledge resources
-
Benchmark observatory · 2026 – present
Leaderboard of Benchmarks
A benchmark of benchmarks. It treats models as probes and compares the evaluation instruments themselves: what capabilities they measure, how their frontiers evolve, how quickly they saturate, how broadly they are covered, and what they cost to run.
-
Living knowledge base · 2026 – present
Scientific Evaluation Environments
An interactive, bilingual map of benchmarks for scientific and engineering AI agents. It explains what each evaluation measures, what its results reveal, and which parts of an agent—skills, scaffolds, data, or post-training—could be improved next.
Small tools
Focused tools that remove friction from research and human-agent collaboration without replacing the projects they plug into.
-
Local writing tool · 2026
PairTeX
A lightweight review layer for existing LaTeX repositories. It renders a local HTML view for comments and in-place review, records feedback as structured files for any coding agent, and leaves the canonical TeX source and Git workflow unchanged.
-
Agent workflow toolkit · 2026
Structured Coding
An installable workflow for planning and executing coding-agent changes one reviewable PR at a time. It connects design agreements, implementation rules, validation evidence, and handoffs so work remains bounded and recoverable across sessions.
Selected research projects
-
XENONnT · Oct 2020 – Feb 2026
Solar-neutrino measurement
Bayesian and maximum-likelihood inference for the first measurement of solar pp neutrinos at ultra-low momentum transfer, including calibration, data selection, and Monte Carlo workflows for uncertainty and sensitivity estimation.
- Corresponding author
- arXiv ↗
-
Nov 2024 – Jun 2025
Adaptive Neural Processes for Extreme Forecasting
Probabilistic forecasting under distribution shift and limited observations, with a focus on few-shot adaptation to extreme regimes never seen during training.
-
Jul 2024 – Oct 2024
Multi-Fidelity Surrogate Modeling for Scientific Simulation
Learning efficient surrogates for computationally expensive, high-variance detector simulations, combining Gaussian Processes and Neural Processes for rare-event emulation.